File size: 10,181 Bytes
08321e5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | """
BenchmarkRunnerV2 — Rigorous evaluation with train/val/test splits,
memory ablation, shuffle control, and contamination detection.
Key difference from V1: BenchmarkRunnerV2 enforces RunMode. In eval_test
mode, no memory is written. This is the only mode whose numbers are trustworthy.
"""
from __future__ import annotations
import json
import logging
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
from purpose_agent.v2_types import RunMode
from purpose_agent.evalport import EvalCase, EvalPort, DictEvalPort, ScoreBundle
from purpose_agent.orchestrator import Orchestrator, TaskResult
from purpose_agent.types import State
logger = logging.getLogger(__name__)
@dataclass
class V2EvalResult:
"""Result of one evaluation case."""
case_id: str
iteration: int
split: str
bundle: ScoreBundle
steps: int = 0
wall_time_s: float = 0.0
@dataclass
class V2BenchmarkResult:
"""Full benchmark result with per-split reporting."""
name: str
results: list[V2EvalResult] = field(default_factory=list)
config: dict[str, Any] = field(default_factory=dict)
started_at: float = field(default_factory=time.time)
finished_at: float = 0.0
def get_split_summary(self, split: str) -> dict[str, float]:
"""Get aggregate metrics for a specific split."""
split_results = [r for r in self.results if r.split == split]
if not split_results:
return {}
n = len(split_results)
pass_rate = sum(1 for r in split_results if r.bundle.passed) / n
avg_steps = sum(r.steps for r in split_results) / n
return {
"n": n,
"pass_rate": round(pass_rate, 3),
"avg_steps": round(avg_steps, 1),
}
def get_improvement_curve(self, split: str = "test") -> list[dict]:
"""Get per-iteration metrics for one split."""
by_iter: dict[int, list[V2EvalResult]] = {}
for r in self.results:
if r.split == split:
by_iter.setdefault(r.iteration, []).append(r)
curve = []
for it in sorted(by_iter):
results = by_iter[it]
n = len(results)
pass_rate = sum(1 for r in results if r.bundle.passed) / n
curve.append({
"iteration": it,
"pass_rate": round(pass_rate, 3),
"n": n,
})
return curve
def summary(self) -> str:
lines = [f"═══ Benchmark: {self.name} ═══"]
for split in ["train", "validation", "test"]:
s = self.get_split_summary(split)
if s:
lines.append(f" {split:>12}: n={s['n']}, pass_rate={s['pass_rate']:.1%}, avg_steps={s['avg_steps']:.1f}")
curve = self.get_improvement_curve("test")
if len(curve) >= 2:
first = curve[0]["pass_rate"]
last = curve[-1]["pass_rate"]
delta = last - first
if abs(delta) < 0.001:
lines.append(f"\n Test improvement: {first:.1%} → {last:.1%} (no significant change)")
else:
lines.append(f"\n Test improvement: {first:.1%} → {last:.1%} ({delta:+.1%})")
return "\n".join(lines)
def save(self, path: str) -> None:
Path(path).parent.mkdir(parents=True, exist_ok=True)
with open(path, "w") as f:
json.dump({
"name": self.name,
"config": self.config,
"splits": {
s: self.get_split_summary(s) for s in ["train", "validation", "test"]
},
"curve": self.get_improvement_curve("test"),
"n_results": len(self.results),
}, f, indent=2)
class BenchmarkRunnerV2:
"""
Rigorous benchmark runner with train/val/test splits and ablation controls.
Key guarantee: eval_test cases NEVER cause memory writes.
Usage:
cases = [
EvalCase(id="t1", input_purpose="...", split="train", ...),
EvalCase(id="t2", input_purpose="...", split="test", ...),
]
runner = BenchmarkRunnerV2(orchestrator=orch)
result = runner.run(cases, train_iterations=3, eval_iterations=1)
print(result.summary())
"""
def __init__(
self,
orchestrator: Orchestrator,
eval_port: EvalPort | None = None,
):
self.orch = orchestrator
self.eval_port = eval_port or DictEvalPort()
def run(
self,
cases: list[EvalCase],
train_iterations: int = 3,
eval_iterations: int = 1,
name: str = "v2_benchmark",
) -> V2BenchmarkResult:
"""
Run benchmark: train split with learning, test split without.
1. Train iterations: run train split cases with RunMode.LEARNING_TRAIN
2. Validation: run validation split with RunMode.LEARNING_VALIDATION
3. Test: run test split with RunMode.EVAL_TEST (no memory writes)
"""
result = V2BenchmarkResult(name=name, config={
"train_iterations": train_iterations,
"eval_iterations": eval_iterations,
})
train_cases = [c for c in cases if c.split == "train"]
val_cases = [c for c in cases if c.split == "validation"]
test_cases = [c for c in cases if c.split == "test"]
# Phase 1: Training
for it in range(1, train_iterations + 1):
logger.info(f"Train iteration {it}/{train_iterations}")
for case in train_cases:
ev = self._run_case(case, it, RunMode.LEARNING_TRAIN)
result.results.append(ev)
# Phase 2: Validation
for case in val_cases:
ev = self._run_case(case, 1, RunMode.LEARNING_VALIDATION)
result.results.append(ev)
# Phase 3: Test (NO MEMORY WRITES)
for it in range(1, eval_iterations + 1):
logger.info(f"Test iteration {it}/{eval_iterations}")
for case in test_cases:
ev = self._run_case(case, it, RunMode.EVAL_TEST)
result.results.append(ev)
result.finished_at = time.time()
return result
def run_cold_warm(
self,
test_cases: list[EvalCase],
train_cases: list[EvalCase],
name: str = "cold_warm",
) -> dict[str, Any]:
"""Compare cold (no memory) vs warm (after training) on the same test set."""
# Cold: eval test cases with empty memory
cold_results = []
for case in test_cases:
ev = self._run_case(case, 0, RunMode.EVAL_TEST)
cold_results.append(ev)
cold_pass = sum(1 for r in cold_results if r.bundle.passed) / max(len(cold_results), 1)
# Train
for case in train_cases:
self._run_case(case, 1, RunMode.LEARNING_TRAIN)
# Warm: eval same test cases after training
warm_results = []
for case in test_cases:
ev = self._run_case(case, 1, RunMode.EVAL_TEST)
warm_results.append(ev)
warm_pass = sum(1 for r in warm_results if r.bundle.passed) / max(len(warm_results), 1)
delta = warm_pass - cold_pass
return {
"cold_pass_rate": round(cold_pass, 3),
"warm_pass_rate": round(warm_pass, 3),
"delta": round(delta, 3),
"improvement_significant": abs(delta) > 0.05,
}
def run_memory_ablation(
self,
test_cases: list[EvalCase],
) -> dict[str, Any]:
"""Run test cases with and without memory to measure memory contribution."""
# With memory
with_results = []
for case in test_cases:
ev = self._run_case(case, 1, RunMode.EVAL_TEST)
with_results.append(ev)
with_pass = sum(1 for r in with_results if r.bundle.passed) / max(len(with_results), 1)
# Without memory (temporarily clear)
saved_lib = list(self.orch.optimizer.heuristic_library)
self.orch.optimizer.heuristic_library = []
self.orch.sync_memory()
without_results = []
for case in test_cases:
ev = self._run_case(case, 1, RunMode.EVAL_TEST)
without_results.append(ev)
without_pass = sum(1 for r in without_results if r.bundle.passed) / max(len(without_results), 1)
# Restore
self.orch.optimizer.heuristic_library = saved_lib
self.orch.sync_memory()
return {
"with_memory_pass_rate": round(with_pass, 3),
"without_memory_pass_rate": round(without_pass, 3),
"memory_contribution": round(with_pass - without_pass, 3),
}
def _run_case(self, case: EvalCase, iteration: int, mode: RunMode) -> V2EvalResult:
"""Run a single case under a specific RunMode."""
start = time.time()
# In EVAL_TEST: save and restore memory state
saved_optimize = self.orch.optimize_every_n_tasks
if mode.is_eval:
self.orch.optimize_every_n_tasks = 999999 # Disable optimization
try:
task_result = self.orch.run_task(
purpose=case.input_purpose,
initial_state=State(data=case.input_state),
max_steps=case.max_steps,
)
except Exception as e:
logger.error(f"Case {case.id} failed: {e}")
task_result = TaskResult(
trajectory=__import__("purpose_agent.types", fromlist=["Trajectory"]).Trajectory(
task_description=case.input_purpose, purpose=case.input_purpose,
),
final_state=State(data={"_error": str(e)}),
)
# Restore
self.orch.optimize_every_n_tasks = saved_optimize
# Evaluate
bundle = self.eval_port.evaluate(
case, task_result.final_state.data, task_result.trajectory,
)
return V2EvalResult(
case_id=case.id,
iteration=iteration,
split=case.split,
bundle=bundle,
steps=task_result.total_steps,
wall_time_s=time.time() - start,
)
|